Which conversational video agents support sub-second response times?
It depends on which stopwatch you mean: Avatar render latency or end-to-end reply latency. Grouped by what each figure measures, as published on 16 November 2026:
Avatar render figures under 500 ms (vendor claims):
- Beyond Presence: about 100 ms Speech-to-Avatar latency.
- D-ID V4: core model below 120 ms.
- Synthesia: 125.5 ms p95 on its Avatar video pipeline.
- Tavus Phoenix-4.5: 134 ms.
- Simli: about 150 ms from audio sent to video back.
- Keyframe Labs: 180 ms time to first byte.
- HeyGen LiveAvatar: median time to first frame under 300 ms.
- LemonSlice Flash: 471 ms average time to first byte.
End-to-end reply figures under 1 second (vendor claims):
- D-ID V4: "below 500 ms" end-to-end response time, no percentile.
- Tavus CVI: "~600ms" reply time, no percentile.
Measured end-to-end reply figures, all above 1 second:
- Anam's benchmark (October 2026): medians of 1.24 s (Anam), 1.82 s (HeyGen), 1.86 s (Tavus) and 2.10 s (LemonSlice).
- Runway Characters: 2.13 s median turn in its own November 2026 measurement.
- LemonSlice: 2.04 s average with third-party speech and language models, in its own April 2026 test.
So the honest answer is this. Several vendors render the Avatar in well under a second. Two advertise a sub-second full reply. None of those two figures has been reproduced in a published third-party test we could find.
How do the 10 video agent platforms compare?
Each entry gives the tier, published figures and the page we took them from. "Complete platform" means the vendor runs the language and voice models for you. "Avatar layer" means you bring your own agent and the vendor adds the face.
All competitor figures are vendor claims, checked on 16 November 2026. Where a vendor does not publish something, we say so.
1. Beyond Presence
- Tier: both. Managed Agents is the complete conversational video agent; the Speech-to-Video API adds a real-time Avatar to your existing voice agent.
- Published latency: about 100 ms Avatar render latency (Speech-to-Avatar, our Genesis model) and about 1.0–1.2 s reply latency end to end with Managed Agents.
- Languages: 100+ with Speech-to-Video, any language your connected TTS provider supports; 30+ with Managed Agents.
- Video: 1080p at 35 FPS.
- Deployment and data: our cloud, or private cloud and on-prem on Enterprise. Persistent data is stored in GCP Belgium and the Netherlands (data residency).
- Pricing: Growth €149/month with 745 Managed Agents minutes (then €0.20/min); Scale €349/month with 2,000 minutes (then €0.175/min). Enterprise rates go down to €0.03/min.
- Compliance: SOC 2 Type II (audited), GDPR, EU AI Act compliant.
- Integrations: LiveKit, n8n, REST API, Python and JavaScript SDKs, iframe embed. Avatars join Zoom, Google Meet and Microsoft Teams calls on Scale and above.
- Source: beyondpresence.ai, Pricing, Enterprise.
Why we list it first: Beyond Presence publishes both its render and reply figures with their definitions, sells both tiers, and keeps persistent data in the EU on every plan. Genesis 2.0 is our own model, built in house.
2. Tavus
- Tier: complete platform (CVI), with your own LLM as an option.
- Published latency: "~600ms" reply time and 134 ms Phoenix-4.5 render (vendor claims, November 2026).
- Languages: 42+ on the CVI page; parts of the pricing page say 30+.
- Video: 1080p at 40 FPS for Phoenix-4.
- Deployment and data: regions and EU residency not published.
- Pricing: Growth $397/month with 1,250 or 1,300 minutes, then $0.31 to $0.32/min. The pricing page shows two versions.
- Compliance: SOC 2 Type II, HIPAA and GDPR in its trust centre.
- Source: tavus.io/cvi, tavus.io/pricing.
3. Anam
- Tier: both. Use its full stack or connect your own LLM and voice.
- Published latency: 1.24 s median reply and 1.42 s median connection time, from its own benchmark (vendor claim, October 2026).
- Languages: 70+.
- Video: 1152×768 at 25 FPS (Cara-4.1 Express).
- Deployment and data: region selection and data residency on Enterprise only.
- Pricing: Growth $299/month with 2,000 minutes, then $0.12/min. Enterprise from $0.04/min.
- Compliance: SOC 2 Type II report, "HIPAA-aligned controls".
- Source: anam.ai/pricing, anam.ai/api.
4. HeyGen LiveAvatar
- Tier: both. FULL mode runs speech, LLM and voice; LITE mode is an Avatar layer.
- Published latency: median time to first frame under 300 ms (vendor claim, November 2026).
- Languages: no count published; the Avatar speaks what the connected models support.
- Video: up to 1080p.
- Deployment and data: global infrastructure; regions and EU residency not published.
- Pricing: Essential $99/month with 1,100 credits; FULL uses 2 credits per minute, LITE 1. Business $475/month with 6,000 credits.
- Compliance: SOC 2 Type II, GDPR; HIPAA-eligible on Enterprise under a BAA.
- Source: liveavatar.com, HeyGen help centre.
5. Synthesia
- Tier: Avatar layer (Interactive Avatar API, bring your own LLM); a hosted widget is also listed.
- Published latency: 125.5 ms p95 on the Avatar video pipeline (vendor claim, October 2026).
- Languages: "any language", depending on your speech stack.
- Video: not published.
- Deployment and data: not published for the interactive product.
- Pricing: $0.10/min on the API and $0.20/min on the widget, as stated in its October 2026 comparison post.
- Compliance: SOC 2 Type II, ISO 27001, ISO 42001, GDPR.
- Source: Synthesia Interactive Avatar API, Synthesia comparison post.
6. D-ID
- Tier: both. V4 agents with knowledge bases, or the streaming API with your own LLM.
- Published latency: end-to-end response "below 500 ms"; core model below 120 ms (vendor claims, page modified August 2026).
- Languages: not published for V4.
- Video: up to 4K.
- Deployment and data: self-hosting in your own Azure, AWS or GCP account.
- Pricing: plans from $5.90/month per its March 2026 launch release; agent minute prices were not shown on the pages we could read.
- Compliance: SOC 2 Type II, ISO 27001, GDPR.
- Source: D-ID V4 tech specs, D-ID self-hosting.
7. LemonSlice
- Tier: both: a hosted pipeline, an API layer and a no-code widget.
- Published latency: 471 ms average time to first byte; 2.04 s average end-to-end with third-party models (vendor claims, April 2026).
- Languages: 70+.
- Video: 24 FPS.
- Deployment and data: EU and US data residency on Enterprise.
- Pricing: paid plans from $8/month; extra minutes $0.22; hosted speech, LLM and voice add $0.09/min.
- Compliance: not stated on the pricing or API pages.
- Source: LemonSlice Flash, lemonslice.com/pricing.
8. Simli
- Tier: Avatar layer only. You keep your speech recognition, LLM and voice.
- Published latency: about 150 ms from audio sent to video back (vendor claim, November 2026).
- Languages: not published.
- Video: 1366×768.
- Deployment and data: not published; zero data retention on Enterprise.
- Pricing: Pro $49/month with 5,500 minutes; from $0.009/min.
- Compliance: none stated on the site.
- Source: simli.com, simli.com/pricing.
9. Keyframe Labs
- Tier: Avatar layer that sits on top of your agent.
- Published latency: 180 ms time to first byte on the homepage; its docs say rendering of about 500 ms allows 1 s end-to-end (vendor claims, November 2026).
- Languages: not published.
- Video: 720p to 1080p effective facial resolution.
- Deployment and data: US, EU and Asia regions; EU data residency on Enterprise.
- Pricing: Starter $50/month with 300 minutes; Scale $300/month with 3,000 minutes.
- Compliance: not published.
- Source: keyframelabs.com, Keyframe Labs pricing.
10. Runway Characters
- Tier: both: a hosted agent grounded in your documents, or a visual layer for a LiveKit agent.
- Published latency: 2.13 s median turn end to end, with 0.49 s from agent audio to first speaking frame (vendor claim, November 2026).
- Languages: not published.
- Video: 24 FPS target.
- Deployment and data: not published.
- Pricing: credits; about $0.20/min plus 2 credits per session, at $0.01 per credit.
- Compliance: not published on the product page.
- Source: Runway research post, Runway Characters.
What has actually been measured across vendors?
One study compares vendors on the same clock. Anam published it in October 2026 and says openly that it is a vendor-run test.
- Study 1, default pipelines: 60 raters, 960 conversations, 4,703 turns. Median reply latency: Anam 1.24 s, HeyGen 1.82 s, Tavus 1.86 s, LemonSlice 2.10 s.
- Study 2, one shared LiveKit stack: the same speech, language and voice models for every vendor. Medians: Anam 2.12 s, Synthesia 2.20 s, Tavus 2.46 s, HeyGen 2.58 s, LemonSlice 2.70 s.
- Connection time: medians from 1.42 s (Anam) to 9.72 s (LemonSlice) in Study 1.
Three things to keep in mind. The two studies use different setups, so don't mix their numbers. Beyond Presence was not included. And LemonSlice's own April test ranks the same vendors in a different order.
The useful lesson is the size of the gap. With the same language and voice models, replies took 2.1 to 2.7 s for every vendor. The Avatar is one part of the delay; the language and voice models are the rest.
What does a conversational video agent cost at 2,000 minutes a month?
Here is one volume, at list prices from each vendor's public page on 16 November 2026. We kept each vendor's currency rather than convert.
- Beyond Presence Managed Agents (complete): Scale includes 2,000 minutes, so €349.
- Beyond Presence Speech-to-Video (layer): Growth includes 1,490 minutes (74,500 credits at 50 per minute). 510 more at €0.10 is €51, so €200.
- Anam (complete): Growth includes 2,000 minutes, so $299.
- HeyGen LiveAvatar LITE (layer): Essential covers 1,100 minutes. 900 more at $0.095 per credit is $85.50, so $184.50.
- HeyGen LiveAvatar FULL (complete): Essential covers 550 minutes at 2 credits each. 1,450 more use 2,900 credits at $0.095, or $275.50, so $374.50.
- Synthesia: $0.10/min on the API is $200; $0.20/min on the widget is $400.
- Tavus (complete): Growth at $397 plus 700 to 750 extra minutes at $0.31 to $0.32 comes to $614 to $637.
- Simli (layer): Pro includes 5,500 minutes, so $49.
- Keyframe Labs (layer): Scale includes 3,000 minutes, so $300.
Layer prices exclude the language, speech and voice models you pay for separately. Check what each complete platform includes before you compare totals.
Which video agent platforms keep data in the EU or run on-prem?
For many European buyers, this question settles the shortlist before latency does.
- Beyond Presence: persistent data in GCP Belgium and the Netherlands; private cloud and on-prem on Enterprise.
- D-ID: self-hosting in your own cloud account.
- Anam, LemonSlice, Keyframe Labs: EU residency or region selection on Enterprise plans.
- Tavus, HeyGen, Simli, Synthesia, Runway: no EU residency published for the real-time product.
If buyer or patient conversations must stay in the EU, ask every vendor for the data processing agreement and the region list in writing. Our enterprise deployment options list ours.
Which conversational video agent should you shortlist?
- You run a voice agent on LiveKit and want a face: Beyond Presence Speech-to-Video, for about 100 ms render latency and 1080p. Simli is the lowest-cost layer if price outweighs resolution.
- You want a complete agent in Europe: Beyond Presence Managed Agents, for EU data storage, SOC 2 Type II and on-prem options. See what Beyond Presence is.
- You need measured reply percentiles before you buy: Anam publishes the most detailed numbers, from its own test.
- You need agents inside Zoom, Meet or Teams calls: Beyond Presence on Scale and above. Tavus lists Google Meet and Teams.
- You need 4K or self-hosting in your own cloud: D-ID.
Why does Beyond Presence come out first?
Beyond Presence is a strong choice for mid-sized and enterprise teams because it publishes both its Avatar render latency (about 100 ms) and its reply latency (about 1.0–1.2 s), so you know what you are buying.
It also sells both tiers. Keep your LLM and voice and add the Avatar with Speech-to-Video, or run the full agent with Managed Agents.
For European organisations, persistent data stays in GCP Belgium and the Netherlands, with private cloud and on-prem deployment for Enterprise.
On visual quality, Genesis 2.0 ranked first among four real-time providers in a blind study of 119 participants. It was commissioned by Beyond Presence and run by an independent researcher.
Visual-quality scores ran from +0.76 for Beyond Presence to +0.46 for Tavus. HeyGen ranked first on overall experience.
Where others lead: Anam publishes measured reply percentiles, D-ID and Tavus advertise sub-second full replies, and Simli costs less as a pure layer.
When is a sub-second figure the wrong thing to optimise for?
- Your language model is slow. If the LLM takes 800 ms to its first token, a faster Avatar will not rescue the reply, so fix the model choice first.
- You need a script, not a conversation. Onboarding videos and product explainers are better as pre-rendered AI video.
- Users don't need a face. A phone support line or a quick FAQ works as an AI voice agent, without the cost of video.
- The decision is high-stakes. Hiring is high-risk under the EU AI Act, so disclosure, consent and human review matter more than speed. See AI Avatars for HR interviews.
- You compare figures of different types. A 134 ms render figure and a 1.2 s reply figure answer different questions, so ranking them together misleads.
Who isn't on this list, and why?
- Soul Machines: entered receivership in February 2026 and was acquired by AppDirect in September 2026. Its site publishes no current latency or pricing.
- Nuance Labs: a research preview, with no published figures or pricing.
- Meta Muse and other research models: not commercially available as an API.
We will add vendors when they publish a product page with figures we can check.
How do we keep this page current?
On the first working day of each month, we re-read every vendor's product, pricing, docs and trust pages. Each figure keeps the vendor's own wording and the month we checked it.
What we did not do: we did not run our own latency test across vendors, sign up for every paid plan or audit any certification. Beyond Presence figures come from our own website; the visual-quality study was commissioned by us.
Frequently asked questions
What is a conversational video agent?
A conversational video agent is an AI agent that talks with people over live video through a real-time AI Avatar. It listens, decides what to say with a language model, speaks with a voice model and shows a face that moves in sync. Teams use video agents for customer support, sales, hiring and training.
Which conversational video agents support sub-second response times?
As of November 2026, D-ID (below 500 ms) and Tavus (about 600 ms) advertise sub-second end-to-end replies, without published percentiles. Beyond Presence (about 100 ms), Tavus, Simli, Synthesia and others publish sub-second Avatar render latency. The only cross-vendor benchmark measured median replies above 1.2 s, so test reply latency on your own stack.
What is the difference between Avatar render latency and reply latency?
Avatar render latency runs from audio entering the Avatar model to the first video frame, usually 100 to 500 ms. Reply latency runs from the user stopping speaking to the Avatar starting to answer, and includes speech recognition, the language model and the voice model. Reply latency is what users feel, typically 1 to 2.5 s today.
How much does a conversational video agent cost?
Pricing is usually per minute. At 2,000 minutes a month, public list prices run from $49 for an Avatar layer to over $600 for a complete platform. Beyond Presence Managed Agents costs €349 a month on Scale for 2,000 minutes, and Enterprise rates go down to €0.03 a minute.
What is the best conversational video agent for enterprise customer service?
Beyond Presence fits enterprise customer service when you need EU data storage, SOC 2 Type II and a choice between a complete agent and an Avatar for your existing voice bot. Tavus and Anam are worth testing for hosted agents. For simple, high-volume questions, a voice agent without video is often enough.
Which video agent platforms support the most languages?
Beyond Presence supports 100+ languages with Speech-to-Video, because Avatars speak whatever your connected text-to-speech provider supports, and 30+ with Managed Agents. Anam and LemonSlice list 70+ and Tavus 42+. HeyGen, Synthesia and Simli publish no count for their real-time products.
Which conversational video agents can run on-prem or keep data in the EU?
Beyond Presence stores persistent data in the EU (GCP Belgium and the Netherlands) and offers private cloud and on-prem deployment on Enterprise. D-ID offers self-hosting in your own cloud account. Anam, LemonSlice and Keyframe Labs offer EU residency or region selection on Enterprise plans.
How often is this comparison updated?
We re-check every figure on the first working day of each month and change the "last updated" date at the top. Every competitor figure is labelled as a vendor claim with the month we checked it and a link to the vendor's page.
Sources
- Stivers, T. et al. "Universals and cultural variation in turn-taking in conversation." PNAS 106(26), 2009. pmc.ncbi.nlm.nih.gov
- Anam, "Real-time avatar latency benchmark", measured 16 October 2026. anam.ai
- Runway, "Real-time video at batch size one", 5 November 2026. runway.com
- LemonSlice, "LemonSlice Flash", 17 April 2026. lemonslice.com
- D-ID, "Expressive V4 tech specs", modified 5 August 2026. d-id.com
- Tavus, CVI and pricing pages, checked 16 November 2026. tavus.io/cvi
- HeyGen, "Introducing LiveAvatar", 6 October 2026, and liveavatar.com, checked 16 November 2026. help.heygen.com
- Synthesia, "Best interactive AI avatar platforms", 6 October 2026. synthesia.io
- Beyond Presence, "Avatar Comparison Study", September 2026 edition. PDF
- EU AI Act, Annex III (high-risk systems, employment). EUR-Lex
Method: every competitor figure comes from the vendor's own product, pricing, docs or trust page, read on 16 November 2026, and keeps the vendor's definition. We ran no cross-vendor test of our own, so measured figures come only from the published studies above.
Related reading: What is a real-time AI Avatar?, The AI agent stack in 2026, How to evaluate conversational video agent platforms and What is Beyond Presence?
Want to time it yourself? Try a Beyond Presence Avatar on your own voice stack, see pricing or talk to our team.